Radiology Systematic Review

Systematic review and evidence synthesis

Use this skill to turn a review question into an auditable evidence-synthesis project, or to test whether an existing review can support its conclusion. It is an independent scientific entry point and a cross-cutting mentor/reviewer module. It does not require a manuscript to exist first.

Core stance

• Protocol first. Freeze the question, eligibility, outcomes, synthesis route, independent unit, subgroup/sensitivity families and deviation policy before result-aware decisions. If work began first, record the protocol as RETROSPECTIVE rather than implying prospective registration.
• Design before pooling. A forest plot is not the objective. Judge clinical, methodological and statistical compatibility before quantitative synthesis; use a structured synthesis without meta-analysis when pooling would erase the question.
• One evidence row is not always one independent study. Detect overlapping cohorts, multiple publications, arms, readers, lesions, images, blocks, sections, cells and spots. Preserve the patient/donor/study hierarchy and never inflate precision through pseudoreplication.
• Risk of bias and applicability are not journal prestige scores. Match the assessment tool to the included design and evaluate domains with evidence anchors. Do not average a fatal domain into a reassuring total score.
• Constructive review. For every problem, state the criterion, consequence, minimum repair, stronger option and cost, closure evidence, and claim that survives if it remains unresolved.
• No invention. Never fabricate a searched database, search date, registration, duplicate screening, included study, extracted value, risk-of-bias judgment, effect size, heterogeneity, certainty rating or review result.

Choose the entry mode

Also declare review_route from the routing table below and the evidence state of supplied artifacts: PLANNED, AUTHOR_REPORTED, PARTLY_VERIFIED, VERIFIED, or NOT_ASSESSABLE.

Required route selection

Open the evidence-synthesis router and select one primary route. Hybrid reviews may activate more than one route, but keep separate eligibility, extraction and synthesis branches when their estimands differ.

Minimal intake

Record available materials and mark missing decision-bearing fields AUTHOR_INPUT_NEEDED:

1. review objective, intended audience/use, scope and preliminary claim;
2. route and question framework: PICO, PIRD, PECO/PICOS, CHARMS-like prediction domains, or a concept-context-population framework for scoping reviews;
3. protocol version/date, registration platform/identifier/status and amendments;
4. databases/platforms, coverage dates, complete strategies, last-search date and grey-literature policy—reported as a search brief until verified by radiology-search; search-strategy peer-review receipt/status and lifecycle state STATIC, LIVING or RETIRED;
5. eligibility, outcomes/estimands, time horizons, effect measure, report/study unit and language/ publication-status restrictions;
6. screening, conflict resolution, exclusion reasons, PRISMA counts and any automation/priority screening contract: tool, model, prompt and version; pilot/calibration evidence; frozen threshold; low-priority exclusion audit and recall safeguard; human override; and final human responsibility;
7. extraction form, reviewer structure, linked reports/overlapping cohorts and usable effect rows;
8. risk-of-bias/applicability tool/version, judgments and evidence anchors;
9. synthesis outputs/code, independent unit, dependence structure, model, heterogeneity, subgroup/sensitivity and certainty evidence.

Required workflow

1. Classify and frame. Choose the route and define the estimand/question before choosing a synthesis method. Do not force every review into PICO; use PIRD for diagnostic accuracy and prediction-specific domains when model development/validation is central.
2. Freeze or reconstruct the protocol. Use the protocol template. Record prospective, retrospective, amended or unregistered status and distinguish dated decisions from later deviations.
3. Create the search handoff. Specify concepts, synonyms, controlled vocabulary needs, eligible designs, databases/platforms, dates and update trigger. radiology-search owns actual database execution, exact query logs, deduplication and corpus acquisition; radiology-citation owns identifier/reference verification. Record whether the complete source-specific strategy was independently peer reviewed, by whom/which role, against which stated criterion, the strategy version reviewed, comments and resolution locators, unresolved issues and whether material amendments require re-review. NOT_REVIEWED is an honest state, not a reason to invent a receipt. Also declare STATIC, LIVING or RETIRED, with update trigger/cadence or retirement reason; a living label does not prove the search is current.
4. Verify selection integrity. Reconcile retrieved, deduplicated, screened, full-text and included counts; document independent screening, conflict resolution and exclusion reasons. For AI-assisted or priority screening, freeze the exact tool/model/prompt/version and threshold before production use; record pilot/calibration corpus and results; preserve every score, recommendation, exclusion and human override; audit a prespecified sample of low-priority exclusions against a recall safeguard; and name the human who retains final inclusion/exclusion responsibility. If the low-priority audit breaches the frozen safeguard, STOP exclusions, expand human screening and recalibrate without deleting the failed pilot. Preserve a study-family identifier linking multiple reports from one cohort.
5. Extract for the question and design. Use the extraction matrix. Separate study/report metadata, population, index/exposure/intervention, comparator/reference standard, outcomes, hierarchy, effect estimates and adjustment from reviewer interpretations.
6. Assess risk of bias and applicability. Open the methods and quality-gates reference, choose a design-matched tool, and use the risk matrix. Preserve domain judgments and supporting anchors; do not invent a numeric quality score.
7. Run the pooling-feasibility gate before meta-analysis. Test compatibility of population, design, target condition/endpoint, time horizon, index/model/intervention/exposure, comparator/ reference standard, effect definition, threshold, unit and dependence. Return POOL, STRATIFY, NARRATIVE, or STOP_FOR_REPAIR with reasons.
8. Specify synthesis. For compatible rows, define effect-size transformation, within-study uncertainty/dependence, random/fixed or hierarchical model candidate, heterogeneity, small-study assessment, prespecified subgroup/meta-regression and sensitivity analyses. Hand computation and statistical model implementation to radiology-stats; never claim an analysis ran from a plan.
9. Handle non-pooling honestly. Group evidence by question-bearing dimensions, tabulate effect direction/magnitude/precision and risk of bias, explain heterogeneity, and state why quantitative synthesis was not justified. Do not vote-count only by statistical significance.
10. Judge certainty and claim ceiling. Apply GRADE or another verified, question-appropriate certainty framework when relevant, or an explicit confidence rationale when it is not. Do not label an informal impression as GRADE. Separate association, prediction, diagnostic performance, intervention effect, reproducibility and mechanism. Cross-sectional or inferred omics concordance does not establish causality.
11. Package reporting and writing handoffs. Send current checklist/version audit and PRISMA/ PRISMA-DTA placement to radiology-reporting; send only fixed, traceable evidence and bounded claims to radiology-writing.
12. Operate a living route when declared. For LIVING, freeze update triggers/cadence, search version, new-record triage, study-family reconciliation, analysis/checklist changes, public version links and retirement criteria. After publication, register corrections, retractions and evidence that can invalidate eligibility, risk-of-bias, synthesis or certainty through the pipeline post-publication lifecycle. LIVING is an operating commitment, not a freshness badge.

Scientific red lines

• Do not change eligibility, outcomes, thresholds, time horizons, subgroup families or pooling decisions because one option yields a favourable conclusion; record justified amendments.
• Do not treat reports from the same cohort, multiple thresholds, time points, outcomes, readers, lesions, sections, cells or spots as statistically independent without modelling or selecting the dependence.
• Do not pool adjusted and unadjusted estimates, development and external-validation performance, per-lesion and per-patient accuracy, incompatible time horizons, or incompatible effect measures merely because they share a label.
• For prevalence or incidence, preserve numerator, denominator or person-time, sampling frame, outcome definition and period. A pooled transformed proportion is not population incidence unless the contributing studies actually estimate incidence on compatible time scales.
• For observational association, preserve the exposure contrast, effect scale, confounder set, adjustment class and temporal order. Correlation or cross-sectional association is not causality, prognosis, diagnostic accuracy or treatment effect.
• Correlation is not agreement. Do not pool unspecified ICC variants, weighted and unweighted kappa, Dice and overlap variants, repeatability coefficients, limits of agreement or reader/model effects as interchangeable. Preserve the measurement object, rater/reader/device hierarchy, repetitions, metric definition/model/form and direction before any synthesis.
• For diagnostic accuracy, do not pool sensitivity and specificity independently while ignoring their threshold relationship when a joint model is required.
• For prediction/radiomics/AI evidence, separate discrimination, calibration and clinical utility; a pooled AUC alone does not establish transportability or patient benefit.
• For bulk/scRNA/spatial/pathology evidence, the donor/patient is normally the inferential unit for patient-level claims. More genes, cells, spots or tiles do not create more biological subjects.
• Mechanistic convergence across modalities strengthens plausibility only to the degree that samples, spatial/temporal scales, perturbations and validation are aligned. Association remains association without intervention, target engagement, rescue or equivalent causal evidence.
• Funnel plots and asymmetry tests are not automatic proof of publication bias and may be uninformative with few or highly heterogeneous studies.
• AI or priority rank is decision support, not an autonomous exclusion authority. Do not silently change a threshold after seeing included studies, discard low-priority records without the frozen audit/recall safeguard, or represent model output as independent duplicate human screening.

Output contract

Return the shortest complete package needed for the selected mode:

1. Review identity — mode, route, objective, framework, scope, protocol/registration state, lifecycle STATIC/LIVING/RETIRED, search-currency state, artifacts and evidence boundary.
2. Question and eligibility lock — population/context, index/exposure/intervention, comparator/ reference standard, outcomes/estimands, horizons, designs, report unit and amendments.
3. Search and selection handoff — executable brief for radiology-search, versioned search-strategy peer-review receipt, dedup/study-family policy, screeners/conflicts, exclusion taxonomy, PRISMA reconciliation and lifecycle update/retirement trigger; when automated, include the frozen tool/model/prompt/version and threshold, pilot/calibration receipt, low-priority audit sample and recall safeguard, human overrides and named final human responsibility.
4. Extraction readiness — required fields, linked reports, missingness, hierarchy, compatible effect rows and unresolved data queries.
5. Risk-of-bias and applicability matrix — design-matched domains, anchors, judgments, consequence and sensitivity role.
6. Pooling-feasibility decision — POOL / STRATIFY / NARRATIVE / STOP_FOR_REPAIR, decisive incompatibilities and eligible synthesis groups.
7. Synthesis specification or results — clearly labelled planned/reported/verified effect measures, model, dependence, heterogeneity, subgroup, sensitivity and small-study assessments.
8. Certainty and claim ceiling — certainty/confidence rationale, what is supported, what is not, residual heterogeneity and applicability limits.
9. Strengths and atomic findings — evidence anchor, severity, criterion, minimum repair, stronger option/cost and closure evidence.
10. Reporting and writing placement — PRISMA flow inputs, Methods, Results, tables/figures, Supplement, Discussion and Abstract/title constraints.
11. Next action — the single next decision-bearing step and all AUTHOR_INPUT_NEEDED fields.

Handoffs and non-ownership

• Database retrieval, exact search execution, deduplication, corpus acquisition and search updates → radiology-search.
• DOI/PMID/reference metadata and claim-support verification → radiology-citation.
• Effect calculation, hierarchical/meta-analytic model fitting, uncertainty, heterogeneity and multiplicity → radiology-stats after this skill supplies compatible rows and hierarchy.
• Current PRISMA/PRISMA-DTA or other reporting-checklist version and item-by-item audit → radiology-reporting.
• Manuscript construction from fixed evidence → radiology-writing; language polish → radiology-polishing.
• Full-manuscript adversarial review → radiology-prereview; final upload-file audit → radiology-submission.
• Published-version monitoring, correction linkage, downstream derivative staleness and retirement coordination → radiology-pipeline; the review owner retains eligibility/synthesis/update decisions.

Do not silently convert a request to find papers into a full review, and do not certify a review as completed when only a protocol, author report or incomplete corpus is available.